* This test does not change the MPI error handling behaviour.
* Hence the default behaviour of abort on exit applies.

* This is intended to help in reproducing issues users are seeing.

* Why the startup of Intel MPI fails in the container running the CI still
  needs investigation.

* Since the Makefile in inst_headers includes a snippet
  generated from the Makefile in src, distclean needs
  to run there first.

* Necessary on FreeBSD and macOS that is.

* It currently does not work in Docker.

* By allowing for a user-pointer to be passed into the constructor,
  thread-safety can be achieved much better.
* Having all the communicators in a read-only struct clarifies
  ownership and simplifies their passing around.

* Use NAG 7.1 in CI setup instead of 7.2 because 7.2 is not yet
  supported in the Docker image.
* Use our libtool patches for autoreconf (required for working NAG builds).

* While the handle remains unchanged and INTENT(IN) suffices
  from the point-of-view of the Fortran compiler,
  having INTENT(INOUT) here makes it much more obvious, that
  the internally referenced object will indeed be changed.

* And rename accordingly.

* xt_mpi_fint_kind is also needed in client code to prevent
  some gfortran warnings.